The performance of the new FEM-package FEAST in connection with the hierachical
solver concept ScaRC as generalised multigrid/domain decomposition approach is based upon 
special data structures requiring meshes consisting of (many) generalized tensor product 
grids (makros). This has impact on the adaptive refinement procedure of such grids, as the 
special grid structure and the underlying data structures do not allow the elementwise 
refinement commonly used in existing adaptive codes. We introduce concepts of corresponding
grid refinement strategies which allow adaptive refinement preserving the high speed of 
FEAST. The refinement strategies rely on three main techniques:

 - adjust the coarse grid level by varying the number of makros
 - refine the grid adaptively in a patchwise manner allowing hanging makro nodes
 - apply local grid deformation if the error source is highly located. 
 
The criteria for mesh refinement stem from error estimation by the DWR (dual weighted 
residual based)-method. We address the problem of the reliability of such estimation by 
numerical examples and present approaches to guarantee reliable goal-oriented error 
estimation.